MNPAT-R: Minnesota's Risk Tool Deserves Transparency, Not Blind Faith

Minnesota built a statewide framework....

Minnesota Judicial Council Policy 524 established a statewide pretrial evaluation framework. The original Minnesota Pretrial Assessment Tool was based on a Hennepin County scale. After a statewide validation study, the Judicial Council approved the revised MNPAT-R in January 2023, and the revised tool took effect on January 1, 2024. [1][2]

Anoka, Cass, Hennepin, Sherburne, and Wright use approved alternative tools rather than the statewide MNPAT-R. That means Minnesota does not operate one identical algorithm in every county. [2]

What the validation study actually found

The original MNPAT produced an AUC of 59.72%, overall accuracy of 58.4%, a false-positive rate of 63.2%, and a false-negative rate of 46%. In plain terms, it was statistically associated with pretrial failure, but its predictive performance was so weak that the study said it could be improved. [3]

Measure	Original MNPAT result
Overall AUC	59.72%
Overall accuracy	58.4%
False-positive rate	63.2%
False-negative rate	46%
AUC - White defendants	61.2%
AUC - Black defendants	51.2%
AUC - Native American defendants	50%
Selected revised model AUC	65.3%

Source: see citation(s) in the surrounding text. AUC is a metric that indicates how well a risk tool can distinguish between someone who will succeed before trial and someone who will fail.

Think of it like this:

• 50% AUC means the tool is no better than flipping a coin.

• 60% AUC means the tool is only slightly better than guessing.

• 100% AUC would mean the tool is perfect.

So, when the MNPAT had an AUC of 59.72%, it means the tool correctly identified the higher-risk person only about 6 times out of 10.

The same report found that the original tool was more predictive for White defendants and that predictiveness was negligible for Black and Native American defendants. The AUC was 61.2% for White defendants, 51.2% for Black defendants, and 50% for Native American defendants. [3] In other words, the program's predictive quality weighs favorably for White defendants.

The false-positive rate was 63.2%

A false positive occurs when the tool identifies someone as likely to fail pretrial even though that person would actually succeed.

The original MNPAT had a false-positive rate of 63.2%.

In simple terms: Among defendants who did not actually fail pretrial, the tool wrongly flagged approximately 63 out of every 100 as potential failures.

These are people who may have appeared in court, followed the rules, and successfully completed the pretrial period, but the assessment still treated them as though they posed a greater risk.

That can lead to unnecessary supervision, electronic monitoring, testing, restrictions, costs, or recommendations for higher bail.

The false-negative rate was 46%

A false negative happens when the tool identifies someone as relatively safe or likely to succeed, but that person later experiences pretrial failure.

The original MNPAT had a false-negative rate of 46%.

In simple terms: Among defendants who actually failed pretrial, the tool failed to identify approximately 46 out of every 100.

So the tool had problems in both directions. It could wrongly label successful people as risky, while also failing to identify nearly half of the people who actually experienced pretrial failure.

That is not simply an imperfect prediction. It is a system that can impose restrictions on the wrong people while failing to recognize risks presented by others.

The original MNPAT was wrong far too often

The original MNPAT had an overall accuracy rate of only 58.4%.

Put plainly: Out of every 100 assessments, the tool was correct about 58 times and wrong about 42 times.

That is not a dependable basis for making decisions that can affect a person’s freedom, release conditions, supervision level, employment, family, and reputation.

The tool’s overall AUC was 59.72%. A score of 50% is equal to random guessing. That means the original MNPAT performed only about 9.72 percentage points better than a coin flip when trying to separate people who would experience pretrial failure from those who would succeed.

The revision improved the tool - it did not make it infallible

The committee replaced the original MNPAT with a simplified three-factor model based on:

·         Whether the defendant is employed or enrolled in school

·         Whether the defendant has another pending criminal case

·         Whether the defendant is already under monitoring or supervision

The revised model produced an AUC of 65.3%, compared with 59.72% for the original tool. That is an improvement of only 5.58 percentage points.

An AUC of 65.3% does not mean the tool is 65.3% accurate in every case. It means that when the tool compares one person who later experiences pretrial failure with one person who succeeds, it assigns the higher risk score to the person who fails about 65 times out of 100.

Put another way: Even after being revised, the tool would still rank the two people incorrectly about 35 times out of every 100 comparisons.

Because random guessing produces an AUC of 50%, the revised MNPAT-R performs only 15.3 percentage points better than chance. That may be useful as a limited piece of information, but it is nowhere near a dependable prediction of what a particular person will actually do.

The validation study did not find the same statistically significant racial difference in the revised model that appeared in the original MNPAT. That is an important improvement, but it does not prove the revised tool is free from bias. [3]

A score cannot perform supervision or recovery

A risk tool can, in theory, organize information and improve consistency. But it cannot contact a defendant, engage an indemnitor, respond to emerging violations, pay a forfeiture, or return a fugitive. It predicts group-level probability; it does not create an accountable third party.

The public-record gap

Minnesota publishes the statewide form and validation report, which is positive. County-specific alternative tools are harder to locate. The public should be able to see each tool's factors, weights, validation population, false-positive and false-negative rates, judicial override rates, outcomes by release type, and every contract paid to administer assessment or supervision.

The proper role

Risk assessment should inform a judge, not replace a judge. It should help identify cases suitable for recognizance and cases needing stronger conditions. It should never be used to erase the constitutional option of sufficient sureties or to present contracted supervision as scientifically guaranteed.

SOURCES & FURTHER READING

[1] Minnesota Judicial Branch, Pretrial Release Initiative — Government / official source.

[2] Minnesota Judicial Branch, MNPAT-R Assessment Tool — Government / official source.

[3] 2023 Minnesota Pretrial Assessment Tool Validation Study — Government / official source.

[4] Minnesota Judicial Branch, MNPAT-R implementation updates — Government / official source.

[5] Minnesota Statutes § 629.74 — Legal authority / official law.

ADDITIONAL READING Harris County Court Watch Data Exposes Failures of Cashless Bail and Unsecured Pretrial Release

Pretrial Justice Institute, “Updated Position on Pretrial Risk Assessment Tools” (2020). The field’s former leading champion concluded that the tools can no longer be part of a fair pretrial system, citing their inability to reliably predict court appearances or new arrests and their reliance on biased data.